Comparative Analysis of Image Segmentation Techniques

Abstract

Image segmentation is a prominent task done in computer vision. Image thresholding is one such technique in image segmentation. Thresholding is a method of categorizing image intensities into two classes and on the basis of that yielding an image which is a binary image, and ideally also has all the fine details of region of interest which an image should have for analysis. Image thresholding is widely used as it reduces the computational cost of processing the image and makes processing feasible in real-world applications like medical imaging, object detection, recognition task, character recognition, etc. This paper dwells into the depth of thresholding techniques to know which technique can perform better on all kinds of images so as to extract region of interest. We found out that not every technique is good for all cases, Otsu’s global thresholding is a promising and faster way to segment and generate a binary image, but works well with images having negligible noise and region of interest already being very much clear in the original image, whereas applying methods like Otsu’s thresholding on sliced blocks of images and then merging them or applying moving averages (sliding windows) on images having noise which is distributed in a specific region of image, moving averages gave result better on images which have distributed gradient noise. Whereas the hybrid technique used are a combination of global and local thresholding.